LLM Agents for Deliberative Collaboration: A Study on Joint Decision Making Under Partial Observability

2025-11-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

A new study investigates large language model (LLM) agents in deliberative collaboration for joint decision-making under partial observability. Researchers formalized this as a cooperative problem with asymmetric information, introducing a scalable benchmark across menu design and task allocation domains. The evaluation, using models like GPT-5.1, DeepSeek-V3.2, and GLM-4.7, revealed that complex deliberative tasks still challenge even advanced LLMs. While top models achieved over 90% normalized reward in simpler menu-numeric tasks, performance significantly dropped in more complex scenarios like task allocation. External mathematical tools generally improved results for models such as DeepSeek-V3.2, GPT-4.1-mini, and Qwen3-32B, but paradoxically reduced performance for others like GPT-5.1 due to agents misinterpreting tool outputs. Diagnostic analysis also showed that the deliberation process itself can provide crucial opportunities for reflection and error correction, sometimes outperforming centralized baselines.

Key takeaway

For AI Engineers designing multi-agent LLM systems for cooperative decision-making, recognize that current LLMs, even advanced ones, struggle with complex deliberative tasks under partial observability. You must prioritize robust agent architectures that facilitate explicit information exchange, reflection, and error correction during deliberation. Do not assume external tools will automatically improve performance; carefully integrate them to avoid "skepticism traps" where agents misinterpret tool outputs, potentially leading to worse outcomes than no-tool baselines.

Key insights

LLM agents face significant challenges in deliberative collaboration under partial observability, yet the deliberation process itself can enhance performance.

Principles

• Partial observability hinders LLM agent collaboration.
• Deliberation provides reflection and error correction.
• External tools require careful integration to prevent misinterpretation.

Method

A multi-module agent framework (observation, planning, decision, conversation) uses chain-of-thought prompting to facilitate deliberative joint decision-making.

In practice

• Formalize cooperative tasks as partially observable joint decision problems.
• Evaluate agent performance using normalized reward and valid ratio.
• Integrate external mathematical solvers for complex optimization tasks.

Topics

• LLM Agents
• Multi-Agent Systems
• Deliberative Collaboration
• Joint Decision Making
• Partial Observability
• Agent Benchmarking
• External Tools

Code references

• wcx21/deliberative-collaboration-agents

Best for: Research Scientist, AI Scientist, AI Engineer

Related on AIssential

• LLM Consortium for Software Design Refinement: A Controlled Experiment on Multi-Agent Collaboration Topologies — Structural adversarial and cross-model review topologies significantly enhance multi-agent LLM software design.
• RetailBench: Benchmarking long horizon reasoning and coherent decision making of LLM agents in realistic retail environments — LLM agents struggle with long-horizon, coherent decision-making in dynamic retail, highlighting a need for better policy consistency.
• Embodied Multi-Agent Coordination by Aligning World Models Through Dialogue — Dialogue in embodied LLM agents reduces action conflicts but can degrade task success without genuine world-model alignment.
• DPBench: Structural Determinants of Multi-Agent LLM Coordination Under Simultaneous Resource Contention — LLMs struggle with simultaneous multi-agent coordination due to "convergent reasoning," even with communication.
• Teaching Machines to Code: A Social Scientist’s Guide to LLM-Based Information Extraction — Multi-agent LLM pipelines excel at complex information extraction by mimicking human collaborative coding processes.
• The cognitive companion: a lightweight parallel monitoring architecture for detecting and recovering from reasoning degradation in LLM agents — A parallel monitoring architecture can detect and mitigate LLM agent reasoning degradation with minimal or zero overhead.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.